Tsetse flies in the time of climate change: What machine learning predicts

Why is the Rekomitjie dataset so special?

A 2020 paper describes the Rekomitjie records as  “one of the most comprehensive longitudinal datasets of tsetse fly count data available to date’’. Apart from the over 200 000 tsetse fly wings, it also contains ovarian dissection data from more than 180 000 female tsetse flies, nutritional data from 40 000 tsetse flies, and catches from more than 10 sampling systems used since 1960. There are also valuable temperature, humidity, and rainfall records, as well as other meteorological data, collected from the same area since 1959.

So, what can the Rekomitjie dataset tell us about what to expect from a warmer climate?

Unlike mammals and birds, insects such as tsetse flies cannot regulate their own body temperature. As such, their rates of development and mortality are strongly influenced by environmental temperatures. Pupae cannot survive at sustained temperatures below 16 °C or above 32 °C.

A 2018 study, based on 27 years of data from Rekomitjie, suggests that temperature increases over the last three decades have already caused major declines in tsetse fly populations in those areas where they traditionally occurred. The study was based on laboratory and field measures of fly densities taken since the 1990s, and nearly continuous records of climatic data collected since 1975 by the Rekomitjie researchers.

Catches of tsetse flies from cattle in Zimbabwe’s Mana Pools National Park have declined from more than 50 flies per animal per catching session in 1990, to fewer than one fly per 10 catching sessions in 2017. Since 1975, mean daily temperatures in the study area have risen by nearly 1 °C. In the hottest month, November, this number is almost double.

Scientists are now trying to understand how increases in temperature may change the distribution and relative abundance of tsetse flies across Zimbabwe and neighbouring countries. Hargrove says the effect of recent and future climate change on the distribution of tsetse flies and other vectors is poorly understood. (Vectors being organisms that transmit a pathogen, disease, or parasite from one animal or plant to another.) “We don’t know, for example, whether the resurgence of malaria in the East African highlands in the 1990s was caused by rising temperatures or by increasing levels of drug resistance and decreasing control efforts. In general, the ways in which climate change will affect the spread of infectious diseases in sub-Saharan Africa are poorly understood because of a lack of empirical evidence.”

He regards the 2018 study mentioned above as one of the first to indicate a link between higher temperatures and an increased susceptibility in certain African regions to sleeping sickness (African trypanosomiasis), transmitted to humans by infected tsetse flies.

Both the Hwange National Park in Zimbabwe and the KNP in South Africa are examples of large conservation areas in which suitable hosts and habitats for tsetse flies abound, Hargrove warns. “Tsetse did occur in these areas in the 19th century, but their numbers were always marginal because the winters were rather too cold. With the massive rinderpest outbreak of the middle 1890s, when the vast majority of ungulates [hoofed animals] died, tsetse disappeared from these areas and have never re-established themselves again. But, if temperatures continue to increase, there is a danger that they may re-emerge.”

To date, however, no one has had sufficient data to model the risks of such an eventuality.

Bringing the Rekomitjie dataset into the 21st century

In 2017, with one of the 27 boxes of valuable tsetse fly data under his arm, Hargrove approached machine learning experts in the Applied Mathematics Division at SU. His objective: automating the measurement of wing shape in tsetse flies.

But why focus on wing shape? To date, researchers have been unable to ascertain whether variations in tsetse flies’ wing shape and size are attributable purely to genetic differences between populations, or whether environmental factors also play a role.

Questions such as this can only be answered on grounds of a large sample size representing several seasons. While SACEMA holds by far the largest collection of tsetse fly wings ever amassed, it is, however, extremely time consuming to manually measure small, fragile things such as insect wings. As a result, previous studies have been limited to sample sizes of a few hundred wings at most.

This is where machine learning comes in. This branch of artificial intelligence enables systems to extract patterns and dependencies from data without being explicitly instructed on how to do so. Applied mathematicians often use it for modelling purposes, which is exactly how the data locked away in those 27 boxes was exploited. The exact process consisted of a number of steps.

Step 1: Photograph the tsetse fly wings with a high-resolution microscope camera. Once the wings have been digitised and numbered, capture the rest of the data associated with each pair in an Excel spreadsheet. This includes the date, time, site, and method of the catch, the name of the dissector, and the results of the dissection, including infection status, estimated age, length, and nutritional status.

To date, Dr Pietro Landi, Prof Cang Hui, and their students in SU’s Department of Mathematical Sciences have photographed about 90 000 pairs of wings from 13 of the 27 boxes.

Step 2: Train a machine learning model to automatically locate 11 anatomical landmarks in any given wing image. For the purposes of machine training data, Prof Willie Brink from the Applied Mathematics Division built a custom annotation tool that allowed him to manually pinpoint landmark positions on 2 000 wing images, all in just a few hours. This initial phase of manual categorisation is vital to the accuracy of the end product: “A model is only as good as the quality of data it is trained on. If I do 1% of the work well, we can train a model to do the remaining 99%,” Brink explains.

To this end, Brink and two of his MSc students, Shane Josias and Mulanga Makhubele, experimented with a number of deep learning architectures.

Step 3: Find postgraduate students to perform a morphometric analysis of tsetse fly wings by quantitatively analysing their form and shape. This would involve studying how the wing shape of tsetse flies changes between males and females, over time, and with the change of seasons.

Then MSc student Dylan Geldenhuys first became involved with the Rekomitjie dataset project when he volunteered to take digital photos of the tsetse fly wings for his graduate student project. This led to a growing interest in the morphometry of the wings — extracting certain features would, he hoped, help machine learning models expose how this morphometry reacts to changes in weather and climate.

Safeguarding the Rekomitjie dataset for the future

“It is crucial that all the data from the Rekomitjie Research Station is archived — with the fullest and clearest possible notation — to ensure that it will be of optimal use to future generations of biologists”, says Hargrove.

So far, the team has barely scratched the surface. Only two of the 27 boxes (14 000 of 205 000 pairs of wings) have been fully digitised.

According to Hargrove, the collection holds enough data to support at least 50 doctoral theses on subjects ranging from birth and ageing processes, physiology, and insect demography to population dynamics, mortality, extinction probabilities, and much more, not to mention the effects of meteorological changes on all of these topics.

“It is fair to say that we have more high-quality field information available on tsetse than anybody has on any other insect species on the planet. This information could, and should, be used to inform studies on other insects.

“If the Rekomitjie data were to be lost, it could never be reproduced. It would be lost, to everybody, for all of time,” says Hargrove.